Data driven approaches to improve understanding of process-based models and decision making

Abstract

This disclosure provides a data-driven and scalable method to discover cause-and-effect relationships in data from natural systems that include sparse data sets. This technique can learn a causal graph from heterogenous data sources by combining embeddings from real data and embeddings from simulated data generated by process-based models. The causal graph is used for what-if analysis in out-of-distribution settings. One application is understanding the factors that affect soil carbon. A causal model created by these techniques can be used to discover cause-and-effect relationships that affect soil carbon. This model has applications such as forecasting soil carbon for a future time point to help inform farm practices. Farm practices, like tilling, may be modified in response to predictions provided by the model.

1 . A computer-implemented method of learning a causal graph from a combination of real data and simulated data comprising:
 obtaining the real data;   generating the simulated data with one or more simulators;   learning a first hidden embedding of the real data using a first neural network;   learning a second hidden embedding of the simulated data using a second neural network;   passing the first hidden embedding through a first multilayer perceptron to the causal graph;   passing the second hidden embedding through a second multilayer perceptron to the causal graph; and   iteratively revising the first hidden embedding, the second hidden embedding, and the causal graph thereby learning the causal graph.   
     
     
         2 . The method of  claim 1 , wherein the real data comprises soil data from an agricultural plot and the simulated data comprises output from the one or more simulators modeling soil processes for the agricultural plot. 
     
     
         3 . The method of  claim 1 , wherein a single data sample includes the real data and the simulated data representing a shared timepoint. 
     
     
         4 . The method of  claim 1 , wherein learning the first hidden embedding comprises learning a first data distribution for the real data and learning the second hidden embedding comprises learning a second data distribution for the simulated data. 
     
     
         5 . The method of  claim 1 , wherein the causal graph is learned such that the likelihood of observing the real data in the causal graph is maximized. 
     
     
         6 . The method of  claim 1 , wherein the first neural network, the second neural network, the first hidden embedding, the second hidden embedding, and the causal graph form a variational auto-encoder based framework. 
     
     
         7 . The method of  claim 1 , wherein iteratively revising the first encoding, the second encoding, and the causal graph comprise node-to-message passing and message-to-node passing. 
     
     
         8 . The method of  claim 1 , further comprising decoding the first hidden embedding and the second hidden embedding with the causal graph to generate an observation that is a final feature vector. 
     
     
         9 . A computer-implemented machine learning model for generating predictions from a combination of real data and simulated data comprising:
 one or more simulators configured to generate the simulated data;   a dataset comprising the real data;   a first trained neural network configured to generate a first hidden embedding from the real data;   a second trained neural network configured to generate a second hidden embedding from the simulated data;   a causal graph learned from the first embedding and the second embedding, wherein the causal graph represents causal relationships between features in the real data and in the simulated data; and   a multilayer perceptron configured to generate a final feature vector from the causal graph, the final feature vector representing a predicted value.   
     
     
         10 . The machine learning model of  claim 9 , wherein the one or more simulators are process-based models that generate the simulated data based on the real data. 
     
     
         11 . The machine learning model of  claim 9 , wherein the real data comprises soil data for an agricultural plot and the predicted value is a soil carbon level. 
     
     
         12 . The machine learning model of  claim 9 , wherein the first trained neural network and the second trained neural network are decoder networks. 
     
     
         13 . The machine learning model of  claim 9 , wherein the causal graph is an encoder network. 
     
     
         14 . The machine learning model of  claim 9 , wherein the causal graph is a directed acyclic graph (DAG). 
     
     
         15 . A method of predicting soil carbon levels comprising:
 obtaining real data for an agricultural plot, the real data comprising soil data and field management practice data;   generating simulated data for the agricultural plot using one or more simulators;   learning a first hidden embedding of the real data;   learning a second hidden embedding of the simulated data;   learning a causal graph from the first encoding and the second encoding; and   generating, by a multilayer perceptron a predicted value for soil carbon.   
     
     
         16 . The method of  claim 15 , wherein the real data and the simulated data are aligned by timepoints. 
     
     
         17 . The method of  claim 15 , wherein the one or more simulators are process-based models. 
     
     
         18 . The method of  claim 15 , wherein the first hidden embedding of the real data and the second hidden embedding of the simulated data are learned using decoder networks. 
     
     
         19 . The method of  claim 15 , wherein the soil data comprises at least one of soil moisture, soil temperature, soil pH, air temperature, wind velocity, greenhouse gasses, solar radiation, sand content, clay content, soil organic matter, the amounts of minerals present, or weather data and the field management practice data comprise data indicating at least one of spreading manure, applying fertilizer, mowing, applying pesticide, or picking up silage. 
     
     
         20 . The method of  claim 15 , further comprising:
 modifying the field management practice data;   generating a second predicted value for soil carbon; and   selecting a field management practice based on the predicted value for soil carbon and on the second predicted value for soil carbon.